sfn_format.md 20 KB

Scalable Screen Font 2.0 File Format

  • File extension: .sfn
  • Mime type: font/x-ssfont

(This spec is about the binary format. See asc_format.md for the human readable ASCII version, which is the "source file" format. This is about the "compiled file" format, created by the converter.)

[[TOC]]

As this font format was primarily designed for screen, it differs from TrueType and OpenType fonts in many aspects. First, it uses pixels, not points (same as when TTF used with 72 DPI). Second, it allows smaller EM sizes and third, it uses screen based Y coordinate direction: the points are stored as (0,0) is in the upper left corner of the grid. Unlike the others, SSFN does not allow glyphs to reach over the grid's edges, all bounding boxes and control points must fit into the grid. So when you render a character into a (let's say) 64 x 64 bitmap, you can be sure that the whole glyph is in it, no parts are missing. Also unlike the others, SSFN fonts can store bitmap and pixel fonts too (even mixed with vector ones, let's say every character has a vector contour, except for the emojis which have a colored pixmap.)

SSFN format does not utilize flexible layout, instead it focuses on the smallest common dominator among font formats, and stores only the minimal required information.

As for character encoding, SSFN exclusively uses UNICODE (ISO-10464) by design. For ligatures, all the standard ligatures (fi, fl, ffi, ae, etc.) and many others have their own UNICODE code points. For those which don't, you can assign code points in the Private Use Area to UTF-8 sequence combinations.

Kerning is splitted into three places in the file: previous character is selected by the kerning fragment descriptor in the Character Mappings table. The next character is selected by a group in the kerning fragment pointed by the descriptor, and finally the coordinate is stored in the Kerning Table pointed by an offset in that kerning group.

All integers are in little-endian format. Offsets are always relative to font magic, except for Kerning Table offsets, which are relative to the table.

The basic structure of the format is as follows:

Block Description
Header A fixed sized struct starting with 'SFN2' magic and basic font information
Strings Human readable UTF-8 strings, like font's name and license
Fragments Sub-glyph information table
Characters UNICODE - glyph mappings
Ligatures UNICODE - UTF-8 ligature mappings (optional)
Kerning Kerning table (character combination distances, optional)
Colormap Palette for color commands and pixmap fragments (optional)
End magic Terminating '2NFS' magic bytes

The description of the detailed on disk format follows.

Header

If the file starts with the bytes 0x1f and 0x8b, then it is stream compressed. You might want to uncompress it first using a gzip decompression filter (inflate). For command line, use gzip -d. The normal renderer is able to uncompress these fonts transparently, but in lack of memory management the simple renderer can't.

An SSFN file can contain more fonts. In this case the file starts with the header

Offset Length Description
0 4 magic, 'SFNC'
4 4 size of the font collection in bytes

And concatenated fonts in SSFN 2.0 format follows (each with its separate header and magic 'SFN2', see the header bellow). With font collections, the entire file is gzip compressed, and not the fonts individually.

When there's only one font in the file, it starts with the 32 bytes long font header:

Offset Length Description
0 4 font magic, 'SFN2'
4 4 size of the font in bytes
8 1 font type (family code and style) (1)
9 1 format revision (2)
10 1 overall width
11 1 overall height
12 1 horizontal baseline (relative to top)
13 1 underline position (relative to top)
14 2 fragments table offset (relative to magic)
16 4 character table offset
20 4 ligature table offset
24 4 kerning table offset
28 4 color map offset (3)

Note(1): type bits are:

BitOffset Description
0 - 3 font family (0-Serif, 1-Sans, 2-Decorative, 3-Monospace, 4-Handwriting)
4 - 7 font style (bit 4: bold, bit 5: italic, bits 6-7: user defined styles)

Note(2): font format revision must be 0 as of writing (implies UNICODE code range 0 .. 0x10FFFF).

Note(3): only if there's a color map. The number of color entires is (font size - 4 - color map offset) / 4.

String Table

Strings are zero terminated UTF-8 strings, up to 255 bytes long each without control characters (less than 32 code points not allowed). If one of the first 6 keys is not defined, then it's coded as an empty string (a single zero byte). If there's no human readable information for them at all, then this block is coded as 6 zero bytes.

String Description
1st Unique name of the font
2nd Font family name
3rd Font sub-family name
4th Version or revision information of the font
5th Manufacturer or designer's name, could be an URL to their website
6th License and copyright information, could be an URL to the license

Other strings (starting from the 7th) are used for ligatures. If they are not defined, there's no zero bytes for them (in other words, only the first 6 strings are mandatory in the string table).

Fragments Table

SSFN does not store glyphs directly. Instead it has fragments which are contours or bitmap blocks with normalized coordinates. A character glyph is constructed by one or more transformed fragments. Usually many glyphs share one or more fragments, this reducing the overall font size considerably.

A fragment starts with a header byte, which most significant bits describe it's type. The fragment descriptors refer to these header bytes by their relative offsets to the magic for fast O(1) access (see section Character Mappings below).

1st byte Description
0xxxxxxx Contour
100xxxxx Bitmap
101xxxxx Pixel map
110xxxxx Kerning group
111xxxxx Hinting information

Contour Fragment

1st byte More bytes Description
00nnnnnn - Contour, up to 64 elements
01NNNNNN 1 Contour, number of elements: (N << 8 + additional byte) + 1

Followed by a byte array for (n+1) commands, where each byte encodes 4 commands (least significant bits first):

Bits Description
0 - 1 1st command
2 - 3 2nd command
4 - 5 3rd command
6 - 7 4th command

The last command byte is padded with zero bits (but only the first 0 command has arguments, not these).

Bits Command Argument bytes
00 move to 2, x y (first two arguments)
01 line to 2, x y
10 quad curve to 4, x y a b (where (a,b) is the control point)
11 cubic curve to 6, x y a b c d

Commands array ((n+1)/4 bytes) is followed by the argument bytes. The first two bytes are reserved for the mandatory first "move to" command (only one "move to" command allowed, other 0 commands are just padding at the end of the commands list with no arguments). Other bytes in the argument list depend on the command's type (number of arguments specified by the table above).

Bitmap Fragment

1st byte More bytes Description
100ppppp h Bitmap dimensions (up to 256 x 256)

First byte encodes type and the pitch (number of bytes per row), the second byte the number of rows (height). Bitmap widths are always multiple of 8.

After the header comes the bitmap data. A set bit means foreground color (part of font face), cleared bits means background (transparency). Least significant bit is on the left, most significant is on the right.

Pixel Map Fragment

1st byte More bytes Description
101SSSSS s,w,h Pixel map with color palette

For pixmaps (w + 1) x (h + 1) compressed indeces follows on ((S << 8) + s + 1) bytes. The index 255 encodes background (transparency). The 254 is foreground and other indeces (0-253) select an ARGB color from the color map and are part of the font face. It is not possible to store separate alpha channel with this format (color map has that). Compressed size is limited to 8192 bytes, and if you want to include a compressed pixmap bigger than 8k, then you're definitely doing the font thing wrong.

Compression: TGA RLE (not that bad ratio and damn easy to decode with a minimal code). Packets start with a header byte, which encodes N (= header & 0x7F). If header bit 7 is set, then next byte should be repeated N+1 times, if not, then the next N+1 bytes are all different indeces.

Kerning Group Fragment

1st byte More bytes Description
110cccNN n Kerning Group Records (up to 1024)

For kerning groups there are ((N << 8) + n + 1) times 8 bytes records: each encoding the first and last code points followed by a Kerning Table offset:

Offset Length Description
0 2 first code point in group
2 1 first code point most significant bits MMMMffff
3 2 last code point in group
5 1 last code point most significant bits OOOOllll
6 2 kerning table offset ((M << 20) + (O << 16) + n) + 1

The code points represents an interval, the next character in the kerning relation. The offset points to a compressed list of coordinates in the Kerning Table. The offset is relative to the table and not to the font magic and describes (last code point - first code point + 1) coordinates.

To determine the kerning for a code point pair, you'll need to look up the first character of the pair in the Character Mappings table (see below). Then you pick the Kerning Group Fragment which includes the next character of the pair, and take the (next character - first code point)th coordinate.

The bits "c" in the header indicates kerning context, for future developments, it is zero for now.

Hinting information

1st byte More bytes Description
111nnnnn - number of autohint grid relative distances (up to 31)

After that cames n relative coordinates, one byte each. The format for horizontal and vertical hinting info are stored alike. If a fragment descriptor with non-zero x points to this fragment, then it is a vertical grid information.

So, for example if the fragment descriptor points to a hinting fragment with x = 9 and n = 3, then the first vertical hinting grid coordinate will be 8 (x-1), the second 8 + first relative distance, the third 8 + first relative distance + second relative distance and finally the last fourth would be 8 + all relative distances. This seems complicated, but guarantees that for example a 2,4,8,10 grid and 5,7,11,13 grid (like for "h", "n", "u") will be stored only once in the file.

Also note that although the format supports hinting grids, the renderer does not need nor use them at the moment.

Character Mappings

SSFN describes the vast UNICODE code point space in runs. There are two different kind of runs, one if there's a glyph defined for the code point, and one if not (skip runs). The table always describes the entire 0 .. 0x10FFFF UNICODE range. If UNICODE Inc. decides to increase that range, then file revision must be incremented.

1st byte More bytes Description
0??????? 5 at least There's a glyph for this character
10nnnnnn - Skip N + 1 code points (up to 64)
11NNNNNN 1 Skip (N << 8 + additional byte) + 1 code points (up to 16128)
11111111 - Skip 65536 code points

The glyph for the 0th character (code point 0 or U+0000) is interpreted as a not defined glyph, which can be shown for every character that is skipped. If no glyph is defined for the 0th character, then the application should render an empty box (it can also write the hex code point inside the box).

The table is a series of: either a skip record, or a defined record with a glyph header and optionally fragment descriptors.

Glyph Header

If the characters table's record describe a glyph at a given code point position, then it starts with a 6 bytes long header (including the aforementioned 1st attribute byte with a zero most significant bit).

Offset Length Description
0 1 attributes, frag type and overlap, 0foooooo
1 1 number of fragments (up to 255)
2 1 glyph width
3 1 glyph height
4 1 advance x (pen movement after the glyph is drawn)
5 1 advance y

Bits in attributes:

  • o is the x overlap in pixels (up to 63, that's 1/4th of total width). Has to be subtracted from pen_x before draw.
  • f encodes large fragment descriptor: 0 = 5 bytes, 1 = 6 bytes.
  • 0 most significant bit must be zero (this is what makes it a "glyph defined" record).

If advance x is zero and y is non-zero, then the character direction is vertical up-to-down.

This header is optionally followed by fragment descriptors. It is possible that a character does not have any fragments (for example the space, U+0020). Descriptors are either 5 or 6 bytes long, depending on "f" bit, but always the same size for one character.

Fragment Descriptors

Depending on the bits in glyph header, for f = 0 the offset is 3 bytes (up to 16M), with f = 1 it is 4 bytes (up to 4G):

Offset Length Description
0 1 x offset
1 1 y offset
2 3/4 fragment offset

The fragment offset is relative to the font magic, and coordinate offsets are relative to character's grid. For hinting grid, the coordinate is one bigger, so x=1 and y=0 encodes vertical grid, and x=0 and y=1 horizontal grid hints (however they both mean 0 offsets to the fragment). For example x=4 and y=0 means vertical grid shifted by 3 to the right. Note that although the format supports hinting grid, the renderer does not need nor use them at the moment.

Similarily, for kerning fragment descriptors x=1 and y=0 encodes vertical kerning groups, while x=0 and y=1 means horizontal groups. Here the actual values of x and y don't matter, they just select the direction.

When x and y offsets are both 255, then instead a fragment offset, a color index follows:

Offset Length Description
0 1 magic 255
1 1 magic 255
2 1 color index
3 2/3 zero, 3 times if "f" set

The color descriptor selects an ARGB from the color map. The next fragment must be rasterized with that color. Color index above 253 are invalid (see Pixmap fragment and Color Map).

There are as many fragment descriptors as the second byte in the glyph header indicates. Their order is important: if grid fragments exists, they must be the first, followed by kerning descriptors if those exists. There can be only one grid and kerning descriptor per direction, meaning up to 4: v-grid, h-grid, v-kern and h-kern. Other descriptors can be repeated as many times as needed. Color commands must always preceed their corresponding contour/bitmap/pixmap fragment descriptors that they colorize.

Ligature Table (optional)

If there's no ligature information associated with the font, then the ligature table offset in the font header is zero. This table assigns UTF-8 sequences to the Private Use Area U+F000 - U+F8FF.

The table is an offset table, each record 2 bytes. These offsets are relative to the font magic, and the last is always a 0 offset. They point to strings in the string table. The first is for the UNICODE U+F000, the second for the U+F001 etc. When a renderer sees one of these sequences in the input stream, it translates more UTF-8 characters into a single UNICODE code point. Although it is not mandatory, but strongly recommended that fonts contain ligatures in lexicographical order to speed up look ups.

Also note that all standard ligatures (like 'fi', 'ffl' etc.) and many more has their own standardized UNICODE code points, no need to assign one from the Private Use Area. Use grep LIGATURE UnicodeData.txt to list them.

Kerning Table (optional)

If there's no kerning information associated with the font, then the kerning table offset in the font header is zero. This table represents the relative coordinates for the kerning.

The table itself is just a concatenated groups of RLE compressed list of signed bytes. Offsets in Kerning Group records point to one of the header byte in one of the RLE packets. The distance of the first and last code point in the fragment specifies how many coordinates to decode in the stream. These are not separated groups, an offset might point in the middle of another group's coordinate list.

Color Map (optional)

Color map only exists if there's at least one pixmap fragment or color descriptor in the font. Its offset is stored in the font header, and the size can be calculated as font size - 4 - color map offset. Color map stores four B,G,R,A bytes for each color (blue is the least significant byte), for index 0-253. Unreferenced colors and values for indeces 254 and 255 are not stored, as those encode foreground and background colors.

End Magic

The SSFN file is closed by the magic in big endian format: '2NFS'.

That's all folks!